Towards Species-Specific Coral Classification in Reef Monitoring Efforts

Abstract

Monitoring the health of coral reefs has traditionally required human labor-intensive effort in the collection and analysis of captured survey data and underwater images. Typical Marine Ecology tasks involve the classification of features within benthic maps and the estimation of coral coverage of the seafloor. In an effort to determine the feasibility of automating part of this process, this work trained and evaluated machine learning models to classify eleven species of stony and fire corals. A binary classifier was developed for each separate species (attaining 95-99% accuracy per model), followed by a single multi-class model (attaining over 92% accuracy). This paper details the architecture, parameterization, and effectiveness of these models as trained on a curated set of images. The models were then evaluated using one square kilometer maps of the seafloor to assess their practicability for automating several image-based analysis tasks on a widespread scale. Developing future monitoring workflows that utilize these machine learning models will minimize the human labor-intensive component of benthic map analysis.